J-PAS: A Neural Network Approach to Single Stellar Population Characterization

TL;DR

This paper tackles the challenge of inferring stellar population parameters from photometric data by leveraging the rich J-PAS 56-band photometry and synthetic SSP libraries. The authors train a neural-network framework on noise-augmented synthetic photometry derived from three SSP libraries (E-MILES, CB19, XSL) to predict SSP age, metallicity, and dust attenuation, demonstrating accuracy comparable to or better than SED-fitting on test sets and robustness against library variations when libraries are combined. They quantify performance as a function of signal-to-noise and show mitigated degeneracies due to independent per-parameter predictions, while also validating the approach with a pilot test on real J-PAS galaxies, which reveals CSP- and emission-line related caveats. Overall, the work establishes a scalable, data-driven path for extracting SSP parameters from upcoming large photometric surveys and highlights directions for incorporating more complex star formation histories and nebular features.

Abstract

J-PAS (Javalambre Physics of the Accelerating Universe Astrophysical Survey) will present a groundbreaking photometric survey covering 8500 deg$^2$ of the visible sky from Javalambre, capturing data in 56 narrow band filters. This survey promises to revolutionize galaxy evolution studies by observing $\sim$10$^8$ galaxies with low spectral resolution. A crucial aspect of this analysis involves predicting stellar population parameters from the observed galaxy photometry. In this study, we combine the exquisite J-PAS photometry with state-of-the-art single stellar population (SSP) libraries to accurately predict stellar age, metallicity, and dust attenuation with a neural network (NN) model. The NN is trained on synthetic J-PAS photometry from different SSP librares (E-MILES, Charlot & Bruzual, XSL), to enhance the robustness of our predictions against individual SSP model variations and limitations. To create mock samples with varying observed magnitudes we add artificial noise in the form of random Gaussian variations within typical observational uncertainties in each band. Our results indicate that the NN can accurately estimate stellar parameters for SSP models without evident degeneracies, surpassing a bayesian SED-fitting method on the same test set. We obtain median bias, scatter and percentage of outliers $μ$ = (0.01 dex, 0.00 dex, 0.00 mag), $σ_{NMAD}$ = (0.23 dex, 0.29 dex, 0.04 mag), f$_{o}$ = (17 %, 24 %, 1 %) at $ i \sim$17 mag for age, metallicity and dust attenuation, respectively. The accuracy of the predictions is highly dependent on the signal-to-noise (S/N) ratio of the photometry, achieving robust predictions up to $i$ $\sim$ 20 mag.

Figures (16)

• Figure 1: Age-metallicity coverage of the three SSP synthesis models used in this work to train the NNs. The grey symbols (dots and histograms) show the combination of the three SSP libraries, while in each panel we show the individual libraries: E-MILES in orange (left panel), CB19 in green (middle panel) and XSL in blue (right panel). Note that the models with ages below < 30 Myr are only included in the CB19 SSP library.
• Figure 2: Examples of synthetic J-PAS photometry from E-MILES SSP for a young (3 Gyr) and an old (10 Gyr) SSP with different metallicities and dust attenuation, as stated in the legend. The fluxes are normalized to the mean value of the flux for each SSP. Note the subtle differences that need to be distinguished for a proper characterization of the SSPs (e.g., compare the red and purple lines).
• Figure 3: Left: average magnitude error of mini-JPAS galaxies for the 57 filters for different observed magnitudes (16, 18, 20 and 22 mag). Right: Synthetic J-PAS photometry obtained after applying random Gaussian variations consistent with the typical errors at different magnitudes. Colors represent different 'mock' observed magnitudes. The thick dashed black line is the original SED.
• Figure 4: Age (left panel), metallicity (middle panel) and dust attenuation (right panel) distribution for the trainning (blue), test (red) and validation (empty black) sub-samples. The sub-samples are randomly selected, showing no differences in the distributions.
• Figure 5: Predicted age vs input age in different magnitude bins. The median $\mu$ and $\sigma_{NMAD}$ of the residuals ($\Delta$ = output - input) are reported for the full sample, and for the three age ranges delimited by the dashed vertical lines. Red points indicate the median values in age bins, with red error bars showing the interquartile range (25th to 75th percentiles) of the distribution. Symbols are color-coded by number density (more populated regions are plotted in yellow). The upper left panel shows the results for all the test sample, while the other three panels are limited to a given $i$ magnitude bin, stated in the legend. The typical S/N in the narrow-band filters is also reported.